AI-Driven Financial Statement Fraud Detection: A Comparative Study with Explainable AI Integration

Purpose: The main objective of this study is to examine whether machine learning classifiers can outperform traditional rule-based statistical fraud detection techniques on valid data, and whether explainable AI can counter the black-box issue.Design/Methodology/Approach: Four Machine Learning (ML) classifiers, including Gradient Boosting, XGBoost, Random Forest, and MLP, were evaluated against four traditional benchmark models using 742 company year-end observations from the SEC EDGAR database (2019-2023), which were labelled based on SEC Accounting and Auditing Enforcement Releases. LIME was applied to make the AI results explainable, and then it was cross validated against forensic accounting indicators.Findings: Gradient Boosting and XGBoost each achieved a ROC-AUC of 0.90, while all traditional results fell below 0.54. LIME explanations consistently identified high leverage, negative ROA and lower liquidity as signs of fraud.Implications/Originality/Value: The findings in this study have direct implications for emerging markets, especially Pakistan’s SECP, as enforcement capacity is limited. A thought on the integration of blockchain is given conceptually, and a framework is proposed. A deployed web application (EMZE) demonstrates pipeline accessibility for non-specialist users.